The website chatbot solved the wrong problem

There is a small ritual most of us perform without thinking. A website loads, a bubble slides up from the corner, and we close it before it finishes typing its greeting. Nobody taught us this. We learned it one canned answer at a time, and the learning is the interesting part: an entire interface pattern managed to train the public to refuse it.

It made sense at the time

It is easy to sneer at the website chatbot now, and mostly unfair. It was born in the economics of support: ticket queues grow, human hours are expensive, and most incoming questions repeat. If software could catch the repeats, people could focus on what remained. Deflection was the honest name for the goal, and teams measured it without embarrassment: conversations contained, tickets avoided, calls that never happened.

The technology shaped the design as much as the goal did. Before software could genuinely read and answer, a conversation had to be assembled in advance: guess the questions, write the answers, wire them into a decision tree. The chatbot was scripted because scripts were all we had. Given those constraints, it was reasonable engineering. The problem is what it was aimed at.

The visitor was never a ticket

The person landing on a business website is usually not a customer with a broken thing. They are a potential customer with a specific situation. Can you take my case? Do you work with companies my size? Does the Tuesday class fit a complete beginner? Their question exists before the relationship does, and it arrives phrased in their words, with their circumstances tangled into it.

To that person, the classic chatbot offered a menu: the business's own predicted questions, arranged as buttons. If your question happened to be on the menu, fine. If not, you were routed in circles or handed the escape hatch, an email form politely renamed "leave a message". A decision tree is not a conversation. It is navigation wearing a chat interface. The chatbot did not replace the sitemap. It read the sitemap aloud.

The over-promise is what stung. A chat window makes a specific claim: someone is here, you can just ask. Every time the reply was a canned paragraph or "I did not understand that", the claim collapsed on the spot.

The wrong problem

Here is the crux. The chatbot was built to solve the business's problem: too many people are contacting us. Within its limits, it did that.

But on a business website, the expensive problem was always the opposite one: too many people are leaving without ever asking anything. The visitor with an unanswered question rarely complains. They close the tab, and nothing records the loss. We made that argument at length in the manifesto: a visitor who leaves with an unanswered question was not unconvinced, they were unattended.

Deflection treats a conversation as a cost to be minimized. On the revenue side of a website, a conversation is not the cost. It is the point. A tool built to end conversations was aimed at precisely the wrong target, and no amount of polish on the widget could fix the aim.

Better technology, same wrong aim

AI removed the constraint that excused the script. Software can now read what a business actually publishes about itself and answer a question nobody predicted, in the visitor's own words. The decision tree is no longer a technical necessity.

But technology does not fix aim. Bolt a language model onto deflection thinking and you get a chatbot with better grammar: still a shield in front of the business, now more fluent, and, if it is allowed to guess, more confidently wrong. The failure of the last decade was never mainly technical. It was the job description.

A different job description

The alternative is not a smarter chatbot. It is a different role.

• A chatbot deflects. A representative attends: success stops being fewer conversations and becomes a visitor who leaves with an answer and a next step, an appointment, an offer, a human's attention.
• A chatbot performs a script. A representative speaks from the business's actual knowledge, the website itself, so its answers are the business's answers and not the script's best guess.
• A chatbot dead-ends at "I did not understand". A representative says "I do not know" plainly and brings a human into the conversation. Notice that this inverts deflection: instead of shielding people from conversations, it hands them exactly the ones worth their time.
• A chatbot is a gadget in the corner of the page. A representative is the website doing its next job: not just explaining the business, representing it.

The distinction is big enough to deserve two names. One is a chatbot. The other is an AI business representative.

The reflex can be unlearned

Reflexes are learned, and they unlearn the same way: one kept promise at a time. The first time a visitor asks a real question in their own words and gets a real answer, the bubble stops being furniture and starts being staff. That will not happen because chat windows get prettier. It happens when the thing behind the bubble changes jobs.

The chatbot solved the wrong problem. The right one is still sitting on almost every business website, unanswered.